Why Screen Covariates Before Imputation?
Imputation is intended to address missing values under a stated rule; it is not a general repair method for every value that looks unusual. A covariate recorded as 178 may be a plausible height in centimeters, an implausible height in meters, or a value whose intended unit is not yet known. If it is treated as missing before its source and unit are checked, an observable data-quality problem can be hidden inside the imputation process.
The order matters in population pharmacokinetic data construction. First determine what was recorded, where it came from, and whether its value and unit are credible in context. Only after that assessment should a value be classified as missing and considered under an imputation rule. Earlier material on Classify Covariate Missingness Before Imputation distinguishes missingness types; here the emphasis is on making sure a suspect observation is not prematurely mislabeled as ordinary missingness.
A Source- and Unit-Aware Screening Sequence
A useful screen proceeds from data meaning to numerical checks, rather than beginning with an arbitrary cutoff. This is especially important because a rare but genuine value may be clinically important, while an apparently ordinary number may have the wrong unit or refer to the wrong visit.
Confirm what the covariate represents, its unit, whether it is baseline or time-varying, and the subject and visit to which it applies. Check the source field definition and any documented transformations.
Look for unit-scale mismatches, impossible codes, misplaced decimal points, sign errors, and confusion between a value and a category code. Standardized units are useful only when the source unit has been identified correctly.
Use source specifications, clinical or measurement context, and relevant within-subject records. A threshold should trigger review when appropriate; it should not automatically delete or replace an observation.
Retain a verified value, correct a confirmed error with documentation, or mark an unresolved suspect observation for review. Classify a value as missing only when it is genuinely unavailable according to the data definition or a documented resolution establishes that no valid value can be used.
Recalculate dependent covariates from verified inputs. Then apply the selected missing-data rule to values that are actually missing, while preserving flags that identify imputed values and their derivatives.
These steps assume that source documentation is available or can be requested, units can be interpreted, and the checks are applied consistently to the relevant records. If the source definition is ambiguous, the correct outcome may be “unresolved,” not an invented correction. A screen can identify records for review without deciding the truth of those records.
Worked Checks
Worked Example: Height, Unit Scale, and a Derived Covariate
Suppose a hypothetical analysis dataset contains a subject’s weight of 71 kg and height of 178, with the analysis specification requiring height in meters. The source record says height was entered in centimeters. Determine whether the analysis value is credible and calculate body mass index (BMI), defined as weight in kilograms divided by height in meters squared.
Check the unit before calculating. The source value 178 cm converts to \(178/100=1.78\) m. The number 178 is plausible as a height in centimeters but not as a height in meters. The mismatch is a unit-scale issue, not evidence that the subject’s height is missing.
Calculate from the converted value.
If 178 were incorrectly treated as meters, the result would be
which is a clear scale warning, not a plausible BMI. The documented source unit supports conversion to 1.78 m. The analysis record should use the standardized value and retain an audit trail of the original value and conversion. If the source unit had not been documented, it would be unsafe to infer centimeters solely because 1.78 m seems likely; the record would need investigation.
Worked Example: An Extreme Laboratory Covariate in Different Units
In a hypothetical study, serum creatinine is recorded as 884 μmol/L for one visit. A review threshold flags values above 300 μmol/L for checking. The laboratory documentation states that the field is serum creatinine in μmol/L. Assess what can and cannot be concluded. For a rough unit conversion, use \(1\ \mathrm{mg/dL}\approx 88.4\ \mathrm{\mu mol/L}\).
The value exceeds the review threshold, so it should be investigated. Converting the value gives
The calculation confirms the scale in another unit; it does not establish that the measurement is erroneous. A high result can reflect a serious clinical state. The appropriate checks include confirming the laboratory field, specimen and visit linkage, the original reported value, and any available source documentation. If those checks confirm 884 μmol/L, retain it as an extreme but observed value and preserve the review flag. If the source instead shows 88.4 μmol/L, the discrepancy is evidence of a possible transcription or scaling error that requires a documented correction.
It would be incorrect to replace the confirmed 884 μmol/L with a typical value, or to label it missing merely because it is extreme. The plausibility threshold identifies a record for review; it does not determine the disposition by itself.
Worked Example: A Longitudinal Covariate That Conflicts With Its Context
A hypothetical participant has recorded body weights of 68.2 kg at baseline, 68.7 kg at week 4, and 687 kg at week 8. The source field is documented in kilograms. Determine an appropriate screening response.
The week-8 value is ten times the scale of the nearby observations and should be flagged for source review. The arithmetic comparison is
That ratio is a useful anomaly signal, but it does not prove a misplaced decimal. Check the original record, entry history if available, and whether the value is linked to the correct subject and visit. If the source confirms 68.7 kg and the analysis field contains 687 kg, correct the analysis value to 68.7 kg with a documented reason, preserving the original value and the correction history. If the source confirms 687 kg, do not silently change it; escalate the discrepancy for clinical or data review. If the source cannot resolve it, classify it as an unresolved suspect value and keep it out of automatic imputation until a decision is made.
The neighboring measurements help prioritize review, but they do not authorize carrying the week-4 value forward. That would substitute an assumed value for an unresolved observation and could conceal a real change or a data error.
Why an Implausible Value Is Not Automatically Missing
An observed value and a missing value carry different information. A confirmed extreme value says that a measurement was recorded and may describe the subject’s condition. A missing value says that no usable observation is available under the data definition. Converting the first into the second without investigation changes the data and may change which records enter an imputation procedure.
The distinction can be expressed as a disposition rather than a single “clean/dirty” decision:
| Finding after review | Appropriate data disposition | Imputation implication |
|---|---|---|
| Value and unit verified; unusual but credible | Retain the observed value and document the review | Do not impute it merely because it is extreme |
| Confirmed source or processing error with a resolvable correct value | Correct with a recorded reason; preserve the original and correction trail | Use the verified correction, not an imputed substitute |
| Value cannot be resolved and is not safe to use | Flag as unresolved; follow a documented analysis disposition | Do not automatically treat it as ordinary missingness |
| No value was recorded or no valid value is available under the data definition | Classify as missing using the applicable missingness convention | Consider the established imputation rule and its assumptions |
This separation also prevents derived covariates from inheriting an unexamined error. For example, BMI calculated from a height with an unresolved unit may be numerically computable but scientifically uninterpretable. Verify the inputs first, recalculate the derived quantity after correction, and preserve the lineage between source values and derived records.
Justification: Why Unit Checks Must Precede Range Checks
A range check compares a value with limits expressed in a particular unit. If the recorded unit is wrong or unknown, the comparison is not meaningful. This follows directly from unit conversion: converting a measurement changes its numerical value by a scale factor while preserving the underlying quantity.
For the same physical height, the numerical values 178 cm and 1.78 m differ by a factor of 100. A rule such as “flag height above 3” has no interpretable meaning until the unit is specified: 3 m is an extreme height, while 3 cm is not a plausible adult height. Thus the order is not merely procedural. A range test applied before unit identification can flag valid values or overlook a scale error, and it cannot establish that a record is missing.
Diagnostics, Alternatives, and Limitations
Useful diagnostics include counts of flags by variable, source, unit, site, visit, and subject; distributions before and after documented corrections; and checks for abrupt within-subject changes. Review whether flags cluster in one source system or unit, since a repeated pattern may indicate a systematic mapping issue rather than many independent biological extremes. Also inspect the values just below and above a chosen threshold: a sharp pile-up can reveal a coding or truncation convention that needs clarification.
These checks have limits. Reference ranges vary with population, clinical context, measurement method, and study design. A fixed cutoff can create false alarms, while a broad cutoff can miss a unit error. A value’s rarity or distance from the sample mean is not by itself evidence of invalidity. Automated rules should therefore generate reviewable flags, not overwrite values.
The main alternatives answer different questions. A hard validity rule can reject values that violate a source-defined constraint, such as a nonnumeric code in a numeric field, but is unsafe when the constraint is only a typical range. Statistical outlier methods identify values unusual relative to the observed dataset, but unusualness does not establish data error and can be especially misleading in heterogeneous populations. Manual source review can resolve specific records, but it requires consistent documentation and may not scale without a triage process. A defensible workflow combines automated flags with source-aware review and leaves unresolved records explicitly unresolved.
Common Pitfalls
- Wrong reasoning: “This value is outside the usual range, so it is missing.” Correction: An extreme observation is still observed. Investigate it and classify it as missing only when the data definition and review support that disposition.
- Wrong reasoning: “The number looks plausible after conversion, so I can assume the source unit.” Correction: A plausible result is not source evidence. Confirm the unit from documentation or treat the value as unresolved.
- Wrong reasoning: “I can replace the suspicious value with the subject’s previous value.” Correction: Carry-forward is not a general correction rule. Resolve the source record and apply the established longitudinal covariate rule only to eligible values.
- Wrong reasoning: “Once I set the value to missing, the imputation procedure will fix the problem.” Correction: Imputation cannot identify whether the original value was a true extreme, unit error, or transcription error. Screening and investigation must come first.
Check Your Understanding
For each question, decide what the screen supports and what additional evidence is needed before changing the analysis value.
- A height of 165 is recorded, but the analysis unit is meters and the source unit is not documented. What is the defensible next step?
- A confirmed serum concentration covariate exceeds a review threshold but matches the laboratory source record. Should it be imputed?
- A value of 52 kg is confirmed in the source, but the analysis dataset contains 520 kg. What disposition is appropriate?
- Why should a derived covariate not be calculated from an unresolved input unit?
- A longitudinal weight changes from 70 kg to 74 kg to 71 kg. Does the return to 71 kg prove that the middle value is erroneous?
Solutions. (1) Flag the record and verify its source unit; do not assume centimeters solely because 165 cm is plausible. (2) No. The value is observed and source-confirmed; retain it with the review documentation unless other evidence changes that conclusion. (3) Correct the analysis value to 52 kg with a documented reason, while preserving the original 520 kg and the correction history. (4) Without a verified unit, the numerical input does not identify a definite physical quantity, so the derived value may be uninterpretable; resolve the unit first, then calculate. (5) No. The pattern may justify review, but it does not establish an error. Check the source record and retain the value if verified.